UJ SN2019 Second Kaggle Competitions

It's my solution for Kaggle Competition - https://www.kaggle.com/c/ujnn2019-2/.

Data Preprocessing and data augmentation

• Raw 10 seconds recording was dividing into 1-second intervals
• Raw audio was transformed to Log-Mel spectrograms
• Only intervals with birds sound were augmented
• Data was augmented online

Model and training

• Resnet with linear layer + linear layers at the end
• Because of unbalanced data (more intervals without birds sound), I used weighted cross-entropy
• For final submission was used ensemble of 15 models